SAR2Opt-Heterogeneous-Dataset

This is an overview of our proposed heterogeneous remote sensing images "SAR2Opt" dataset.

SAR2Opt dataset can be used as a benchmark in change detection and image translation on remote sensing images.

Description of dataset

We manually selected ground points on each pair of SAR-optical images to perform fine registration. All the pathces were cropped in size of 600*600 pixels after registration. If you want to use a pre-trained model, you could resize the patches first.

Dataset

You can get the dataset from:

• Google Drive
• BaiduDisk (hy8d)

Image-to-image translation results on SAR2Opt dataset

Here are some translated results on our SAR2Opt dataset with well-known GAN-based methods, which have been included in our GRSL paper.

Usage of codes

We slightly modofied the original codes to make our dataset suitable for the models. And we are grateful to the authors who have shared their codes kindly.

Pix2Pix

• train Pix2Pix:

python train.py --dataroot ./datasets/combined_sar2opt --name sar2opt_pix2pix --model pix2pix --direction AtoB

• test Pix2Pix:

python test.py --dataroot ./datasets/combined_sar2opt --name sar2opt_pix2pix --model pix2pix --direction AtoB

BicycleGAN

• train BicycleGAN:

bash ./scripts/train_scripts/train_combined_sar2opt.sh

• test BicycleGAN:

bash ./scripts/test_scripts/test_combined_sar2opt.sh

CycleGAN

• train CycleGAN:

python test.py --dataroot ./datasets/sar2opt --name sar2opt_cyclegan --model cycle_gan

• test CycleGAN:

python test.py --dataroot ./datasets/sar2opt --name sar2opt_cyclegan --model cycle_gan

MUNIT

• train MUNIT:

python train.py --config configs/sar2opt_folder.yaml

• test MUNIT:

python test.py --config configs/sar2opt_folder.yaml \
               --input_folder datasets/sar2opt/testA/ \
               --output_folder results/sar2opt/images/ \
               --checkpoint outputs/sar2opt_folder/checkpoint/YOUR_MODEL.pt \
               --a2b 1 

NICE-GAN

• train NICE-GAN:

python main.py --dataset sar2opt

• test NICE-GAN:

python main.py --dataset sar2opt --phase test

CUT

• train CUT:

python train.py --dataroot ./datasets/sar2opt --name sar2opt_CUT --CUT_mode CUT

• test CUT:

 python test.py --dataroot ./datasets/sar2opt --name sar2opt_CUT --model cut

Attn-CycleGAN (ASGIT)

• train BicycleGAN:

bash my_scripts/train_asgit_v2.sh

• test BicycleGAN:

bash my_scripts/test_asgit_v2.sh

Baselines we referenced here are:

• Pix2Pix
• BicycleGAN
• CycleGAN
• MUNIT
• NICE-GAN
• CUT
• Attn-CycleGAN

Citation

If you find this dataset valuable in your projects, please cite our paper below:

@article{zhao2022comparative,
  title={A Comparative Analysis of GAN-based Methods for SAR-to-Optical Image Translation},
  author={Zhao, Yitao and Celik, Turgay and Liu, Nanqing and Li, Heng-Chao},
  journal={IEEE Geoscience and Remote Sensing Letters},
  year={2022},
  publisher={IEEE}
}